Resolving the CN-MCI Boundary: A ResoNet Framework for Dynamic Multimodal Alzheimer’s Classification

Karen Kuo, Wei-Yu Chen, Jenhui Chen

Keywords: Alzheimer's Disease, Mild Cognitive Impairment (MCI), Neuroimaging, Multimodal Fusion, Feature-wise Linear Modulation (FiLM), Boundary-Aware Learning, Deep Learning, Classification, 3D MRI, Supervised Contrastive Learning, ResoNet

TL;DR: ResoNet dynamically modulates 3D MRI features with patient-specific tabular data to resolve the ambiguous CN-MCI boundary in Alzheimer's diagnosis.

Abstract: Distinguishing mild cognitive impairment (MCI) from cognitively normal (CN) subjects remains a critical challenge, as most frameworks fail by treating tabular data as static covariates rather than dynamic modulators of MRI features, and by ignoring the ambiguous CN-MCI decision boundary. We propose ResoNet (Resolution Network), a novel multi- modal framework that tackles these gaps. Critically, ResoNet achieves 93.55% accuracy on the difficult CN vs. MCI sub-task and 92.75% overall three-class (CN/MCI/AD) accuracy on the ADNI dataset. This effectiveness is driven by four core contributions:(1) A Cross- modal Feature-wise Linear Modulation (FiLM) Gating mechanism, where patient-specific tabular data generates parameters (γ,β) to dynamically modulate 3D MRI feature maps, creating personalized representations.(2) A joint boundary-aware objective combining a Supervised Contrastive (SupCon) Loss.(3) A dedicated Auxiliary CN-MCI Binary Head, which forces the model to learn finer discriminative features for the most ambiguous boundary.(4) A Class-aware Multimodal MixUp augmentation strategy that selectively increases mixing probability for MCI-involved pairs, further regularizing the boundary. ResoNet significantly improves diagnostic accuracy and model robustness, making it a powerful tool for resolving early-stage AD diagnostic ambiguity.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Learning with Noisy Labels and Limited Data

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 18